You don't have to tolerate agent/AI going haywire. In a simple example, say of multiple parallel generations. It's compute intensive and it reduces the probabi

You don't have to tolerate agent/AI going haywire. In a simple example, say of multiple parallel generations. It's compute intensive and it reduces the probability of your agent going haywire. You need mechanisms and evals to detect the best output in this scenario of course, that is still important. With more compute, you are preventing your final output to be haywire despite the variance.